ThursdAI - The top AI news from the past week

Daniel Han, open-source ML engineer and Onslaught contributor who builds tools for fine-tuning and quantization, joins to unpack model compression and deployment. He explains weight quantization, KV cache trade-offs, and real-world limits of TurboQuant. The conversation also reacts to Gemini 3.1 Flash Live, new voice and music models, and key infra and quota dramas in the AI world.

Yam Peleg, an AI practitioner who probes tooling and enterprise implications, and Wolfram Ravenwolf, an engineer who runs WolfBench performance analyses, dig into Anthropic's 1M-context Opus, OpenAI's GPT‑5.4 Mini and Nano, Mistral's new small model, Minimax 2.7's self‑improvement loops, Jensen's OpenClaw remarks, Composer 2 and benchmark showdowns. Short takes on agentic tooling, benchmarks, and model tradeoffs.

Dota, an open-source developer behind Paperclip.ing, explains agent orchestration and heartbeat-driven 'zero-human' workflows. Chris Alexiuk, a NVIDIA product research engineer, outlines NemoTron 3 Super and NVIDIA’s open-source commitments. They discuss AutoResearcher-style autonomous research, the fruit fly brain upload, OpenClaw’s viral adoption, and NVIDIA’s $26B open-source push.

Big model news with a surprise GPT-5.4 drop that adds 1M-context, faster modes, and mid-thought steering. Tense industry politics around Anthropic and a government supply-chain designation. Open-source drama as a prominent researcher departs Qwen and Alibaba responds. New tools and benchmarks like WolfBench and Qwen 3.5 small spark debate about performance and running models locally.

Nader Dabit, a developer-relations/product practitioner at Cognition who worked on Devin, and Ben Broca, engineer and founder of Polsia, a platform for autonomous AI companies. They discuss agentic tooling, Devin 2.2 features for engineering automation, and how Polsia runs and scales autonomous businesses. Rapid autonomy, tool ecosystems, and the race to run AI companies 24/7 are the main themes.

Ryan Carson, startup founder and writer, shares his CodeFactory workflow and production tips. Nisten Tahiraj, AI practitioner, runs vibe checks on models like Qwen 3.5 and demos front-end experiments. Wolfram Ravenwolf, engineer and evaluator, breaks down TerminalBench and benchmarking quirks. They debate Gemini 3.1 Pro, Sonnet 4.6, model harness effects, and agent-driven code pipelines in short, lively segments.

Olive Song, senior reinforcement learning researcher at MiniMax who led the M2.5 release and the Forge RL framework, joins to explain MiniMax M2.5’s 80.2% SWE‑Bench result and compact 10B active‑param design. Conversation covers RL training strategies, speed and efficiency optimizations, open‑source tradeoffs, and how agentic workflows speed iteration. Quick, technical and forward‑looking.

Vibi (VB), Developer Experience Lead at OpenAI, shares a brief bio as a dev tools and product lead. They walk through the new Codex app, its automations, skills marketplace and work trees. The conversation covers GPT-5.3-Codex improvements like speed, token efficiency and steerability. The discussion also touches on developer UX choices and the Frontier agent marketplace.

Wolfram Ravenwolf, an AI engineer known for technical commentary on model releases, joins to unpack this whirlwind week. They cover Genie 3’s real-time persistent worlds and Chrome’s new agentic browsing for billions. Open-source breakthroughs like Kimi K2.5, Trinity Large, Grok Imagine’s API and advances in video and image models get technical analysis and reactions.

Dan Pagan, a tech entrepreneur and early adopter of CloudBot, joins to explore the innovative personal AI assistant that can improve itself and manage tasks through phone control. He shares insights on CloudBot's architecture, skills setup, and local security implications. The discussion highlights its impressive capabilities like code writing and life organization. Additionally, the team delves into GLM 4.7 Flash's advancements and Anthropic's newly released constitution, which sets ethical standards for AI behavior.